NOTICE: this is the authors version of a work that was accepted for

publication in Information Sciences. Changes resulting from the

publishing process, such as peer review, editing, corrections, structural
formatting, and other quality control mechanisms may not be reflected in
this document. Changes may have been made to this work since it was
submitted for publication. A definitive version was subsequently
published in Information Sciences, Vol.181, no.9 (May 2011).
DOI: 10.1016/j.ins.2011.01.006

Polynomial modeling for time-varying systems based on a particle
swarm optimization algorithm
1
*
Kit Yan Chan,
1
Dillon S. Tharam

and
2
Che Kit Kwong
1
Digital Ecosystem and Business Intelligence Institute, Curtin University of Technology, WA 6102,
Australia
2
Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom,
Hong Kong
Abstract In this paper, an effective particle swarm optimization (PSO) is proposed for
polynomial models for time varying systems. The basic operations of the proposed PSO
are similar to those of the classical PSO except that elements of particles represent
arithmetic operations and variables of time-varying models. The performance of the
proposed PSO is evaluated by polynomial modeling based on various sets of time-
invariant and time-varying data. Results of polynomial modeling in time-varying systems
show that the proposed PSO outperforms commonly used modeling methods which have
been developed for solving dynamic optimization problems including genetic
programming (GP) and dynamic GP. An analysis of the diversity of individuals of
populations in the proposed PSO and GP reveals why the proposed PSO obtains better
results than those obtained by GP.
Keywords Particle swarm optimization, time-varying systems, polynomial modeling,
genetic programming




*
Kit Yan Chan is the corresponding author. His email address is Kit.Chan@curtin.edu.au
1. Introduction
Genetic programming (GP) [25, 26] is a commonly used evolutionary computation
method which is used to generate polynomial models for various systems such as
chemical plants [38], time series systems [21], nonlinear dynamic systems [56], object
classification systems [1, 65], machine learning systems [27], feature selection systems
[43], object detection systems [37], speech recognition systems [11], control systems [5]
and mechatronic systems [61]. The GP starts by creating a random initial population of
individuals, each of which represents the structure of a polynomial model. Evolution of
individuals takes place by mutation and crossover over generations, and individuals with
high goodness-of-fit are selected as survivors in the next generation. The evolutionary
process continues until the diversity of individuals of a population saturates to a low level
or no progress can be found.
Observations reveal that polynomial models represented by individuals in the GP
are distinct from each other in early generations. As the GP is progressing, polynomial
models represented by individuals converge to a form, which achieves relatively higher
goodness-of-fit in the population. Vaessens et al. [59] and Reeves [55] put this
population-based optimization method into the context of local searches. Maintaining
population diversity in GP is a key to preventing premature convergence and stagnation
in local optima [17, 40]. Using GP, it is difficult to develop optimal polynomial models
for time-varying systems whose structures or coefficients vary over time while the
diversity of individuals in a population is low. Time-varying characteristics can
commonly be found in many industrial systems [6, 22, 41, 44, 57, 66, 36, 34]. To develop
models for time-varying environments, Wagner et al. [60] developed a GP approach in
which a varying window for capturing significant time series is proposed to generate time
series models based on time series data. This approach cannot be applied for generating
models for time-varying systems if the nature of the data is not all in time series formats.
While mechanisms implemented on evolutionary algorithms have been well studied for
solving various dynamic optimization problems [64], those implemented in GP have not
been thoroughly studied for the development of polynomial models in time-varying
environments. It is essential that an effective algorithm be developed for generating
models that deal with time-varying characteristics, given their occurrence in many
industrial systems.
Another more recent population based optimization method, particle swarm
optimization (PSO) [15], inspires the movements of a population of individuals seeking
optimal solutions. The movement of each individual is based on its best position recorded
so far from previous generations and the position of the best individual among all the
individuals [28, 29]. The diversity of the individuals can be maintained by selecting PSO
parameters which provide a balance between global exploration, based on the position of
the best individual in the swarm, and local exploration based on each individuals best
previous position. Each individual can move gradually toward both its best position
recorded to date and the position of the best individual in the population. Kennedy and
Eberhart [29] demonstrated that PSO can solve many difficult optimization problems
with satisfactory results. PSO outperforms evolutionary computation methods for solving
various static optimization problems [13, 31, 53, 62], and various dynamic optimization
problems [2, 9, 8, 52] in which the optima or landscapes of the problems vary over time.
Although PSO can obtain satisfactory results when solving various dynamic optimization
problems, PSO has not currently been used on polynomial modelling for time-varying
systems. The development of PSO for polynomial modelling for systems with time-
varying characteristics is a new research area.
In this paper, a PSO is proposed for the development of polynomial models for
time-varying systems in which the system coefficients vary over time. The basic
operations of the proposed PSO are identical to those of the classical PSO [12] except
that the elements of individuals are represented by arithmetic operations and system
variables of polynomial models. The representation of elements takes the form of
grammatical swarm [47, 48] or grammatical evolution [46]. The performance of the
proposed PSO in the present paper is evaluated by developing models based on several
sets of time-varying data which are generated based on time-varying functions with
different time varying characteristics. In order to provide a comprehensive evaluation, a
comparison is conducted of the results obtained by the proposed PSO with:
(a) classical GP [46] - in which the representation of individuals of population is
identical to the one used in the proposed PSO;
(b) dynamic GP - which is integrated with a recent mechanism [63] for solving
dynamic optimization problems;
(c) dynamic PSO - which is integrated with a recent mechanism [2] for dynamic
optimization problems.
Even if additional computational effort is used in the dynamic PSO to maintain
the diversity of individuals, no significant difference in diversity can be found between
the proposed PSO and the dynamic PSO. Compared with the two GPs, the results indicate
that the proposed PSO outperforms both classical GP and dynamic GP in developing
polynomial models for systems with both time-invariant data and time-varying data. The
results can be explained by the diversity of individuals in the proposed PSO, which can
be maintained in both early and later generations. The individuals of the proposed PSO
continue to explore the solution spaces over the generations. In contrast the individuals of
both the GP methods start to converge and get stuck on a solution after early generations.
This paper is organized as follows. Section 2 presents the operations of the
proposed PSO. The experimental set-up for testing the proposed PSO, and the data sets
used for evaluating the proposed PSO are presented in Sections 3.1 and 3.2 respectively.
The experimental results and the analysis of the experimental results are presented in
Section 3.3 and 3.4 respectively. Finally, conclusions and suggestions for further work
are given in Section 4.

2. Particle swarm optimization
A time-varying system can be formulated as follows:
y = f
t
(x
1
,x
2
,x
m
) (1)
where y is the output response, x
j
, j=1,2,m, is the j-th variable of the time-varying
system, and f
t
is the functional relationship of the time-varying system at time t. Based
on a set of data which represents relations between the output response y and the
variables, x
1
, x
2
, ..., x
m
at time t, the time-varying system
t
f in (1) can be generated in a
polynomial form with constant coefficients at time t. The data set at time t is defined by
( ) ( ) ( ) ( ) { }
D
N
i
D D
i t y i t t
1
, , ,
=
= x D , where the corresponding values of the i-th data at time t is
( ) ( ) ( ) ( ) ( )
m D
m
D D D
R i t x i t x i t x i t e = , ,..., , , , ,
2 1
x and the corresponding value of the response
output of the i-th data at time t is ( ) R i t y
D
e , . Where ( ) t D is available,
t
f can be
generated as the high-order high-dimensional Kolmogorov-Gabor polynomial in
expression (2):
( ) ( ) ( ) ( ) ( )
[
= = = = = = =
+ + + + =
m
i
i m
m
i
m
i
m
i
i i i i i i
m
i
m
i
i i i i
m
i
i i
m
m
x t a x x x t a x x t a x t a t a y
1
... 123
1 1 1 1 1 1
0
...
1 2 3
3 2 1 3 2 1
1 2
2 1 2 1
1
1 1

(2)
where ( ) t a
0
, ( ) t a
1
, ( ) t a
2
, ...., ( ) t a
m
, ( ) t a
11
, ( ) t a
12
, ..., ( ) t a
mm
,... and ( ) t a
m mm...
are the
polynomial coefficients at time t. Equation (2) is a universal format of the polynomial
model if the number of terms in equation (2) is large enough [18]. In this paper, a PSO is
proposed in order to generate the time-varying model at time t based on equation (2),
using an available set of data at time t. Based on [12], the proposed PSO uses a number of
individuals, which constitute a swarm, and each individual represents a time-varying
model. Each individual traverses the search space to trace the polynomial model of the
time-varying system whose system coefficients vary over time.
In the PSO, each individual is represented by the system variables (x
1
, x
2
, , and
x
m
) and the arithmetic operations (+, - and *) of the system model as defined in (2).
m is the number of variables of the system model. A similar mechanism was first
proposed by Kennedy and Eberhart [30] for representing discrete binary variables, and
has been applied to the PSO for solving flowshop scheduling problems [31, 51, 58]. The
i-th individual at generation g is defined as
( ) ,1 ,2 ,
, ,...,
p
g g g g
i i i i N
P p p p = ; where m N
p
> ;
pop
N i ,..., 2 , 1 = ;
pop
N is the number of individuals of the swarm;
p
N
is the number of
elements of the individual; and
pop
N is an odd number;
,
g
i k
p is the k-th element of the i-th
individual at the g-th generation, and
,
g
i k
p is in the range between 0 to 1 i.e. { }
,
0...1
g
i k
p e .
If the value of
p
N
is large, a larger number of terms can be generated in the model, and
the model can better fit the data which is used for model development. However, a model
may contain too many unnecessary and complex terms. A complex, over-parameterized
model with a large number of parametric terms reduces the transparency and ease of
interpretation of the model leading to overfitting problems. To prevent the PSO from
generating models which are too complex, the value of
p
N
has to be selected carefully.
The value of
p
N
can be determined based on the trial and error method, and the value of
p
N
cannot be set too high, otherwise redundant terms can be produced. If the number of
variables of the system model is 4,
p
N
can be initially set as 10. If the modelling error
obtained by the PSO is not satisfactory, the value of
p
N
can be increased until a
satisfactory modelling error is achieved. If the modelling error obtained by the PSO is
satisfactory, the value of
p
N
can be decreased until just before an unsatisfactory
modelling error is achieved.
The elements in odd numbers (i.e.
,1 ,3 ,5
, , ,...
g g g
i i i
p p p ) are used to represent the
system variables, and the elements in even numbers (i.e.
,2 ,4 ,6
, , ,...
g g g
i i i
p p p ) are used to
represent the arithmetic operations. For odd k, if
,
1
0
1
g
i k
p
m
< s
+
, no system variable is
represented by the element
,
g
i k
p . If
,
1

1 1
g
i k
l l
p
m m
+
< s
+ +
with l>0,
,
g
i k
p represents the l-th
system variable,
l
x . System variables represented by the individual are summarized in
Table 1.

Table 1: Representation of system variables
The k-
th
element
,
1
0
1
g
i k
p
m
< s
+

,
1 2

1 1
g
i k
p
m m
< s
+ +
,
2 3

1 1
g
i k
p
m m
< s
+ +

.
,
1
1
g
i k
m
p
m
< s
+
The
system
variable
No system
variable
1
x
2
x .
m
x
*
k is an odd number
In the polynomial model, +, - and * are the only three arithmetic operations
considered. For even k, if
,
1
0
3
g
i k
p < s ,
,
1 2

3 3
g
i k
p < s and
,
2
1
3
g
i k
p < s , the element
,
g
i k
p
represents the arithmetic operations +, - and * respectively. Arithmetic operations
represented by the individual are summarized in Table 2.
Table 2: Representation of arithmetic operations
The k-th
element
,
1
0
3
g
i k
p < s
,
1 2

3 3
g
i k
p < s
,
2
1
3
g
i k
p < s
The arithmetic
operations
+ - *
*
k is an even number
For example, the i-th individual at generation g with 11 elements is used to
represent a polynomial model of the time-varying system at time t, which consists of 4
system variables (i.e. x
1
, x
2
, x
3
and x
4
):
g
i
p
1 ,

g
i
p
2 ,

g
i
p
3 ,

g
i
p
4 ,

g
i
p
5 ,

g
i
p
6 ,

g
i
p
7 ,

g
i
p
8 ,

g
i
p
9 ,

g
i
p
10 ,

g
i
p
11 ,

0.18 0.41 0.94 0.92 0.41 0.89 0.06 0.35 0.81 0.01 0.74
The elements in the individual are within the following ranges:
g
i
p
1 ,

g
i
p
2 ,

g
i
p
3 ,

g
i
p
4 ,

g
i
p
5 ,

g
i
p
6 ,

g
i
p
7 ,

g
i
p
8 ,

g
i
p
9 ,

g
i
p
10 ,

g
i
p
11 ,

0<
0.18
1/5
1/3<
0.41
2/3
4/5<
0.94
5/5
2/3<
0.92
3/3
2/5<
0.41
3/5
0<
0.08
1/3
0<
0.06
1/5
0<
0.35
1/3
4/5<
0.81
5/5
0<
0.01
1/3
3/5<
0.74
4/5


Therefore, the model is represented in the following form:
g
i
p
1 ,

g
i
p
2 ,

g
i
p
3 ,

g
i
p
4 ,

g
i
p
5 ,

g
i
p
6 ,

g
i
p
7 ,

g
i
p
8 ,

g
i
p
9 ,

g
i
p
10 ,

g
i
p
11 ,

0 -
4
x *
2
x + 0 +
4
x +
3
x

which is equivalent to:
( )
3 4 2 4
0 0 x x x x
g
i
+ + + = x
or
( )
3 4 2 4
x x x x
g
i
+ + = x .
The PSO is used only to find the structure of the polynomial and not the
coefficients. The system coefficients a
0
(t), a
1
(t), a
2
(t) and a
3
(t) are determined after the
structure of the time-varying model at time t is established, where the number of
coefficients is 4. The completed time-varying model at time t is represented as follows:
( ) = x
g
i
f a
0
(t) a
1
(t)x
4
x
2
+ a
2
(t)x
4
+ a
3
(t)x
3
In this research, the system coefficients a
0
(t), a
1
(t), a
2
(t) and a
3
(t) are determined
by the orthogonal least squares algorithm (OLSA) [2, 5], which has been demonstrated to
be effective in determining system coefficients in polynomial models [39]. Details of the
orthogonal least squares algorithm can be found in [3, 7].
The polynomial model represented by each individual is evaluated based on the
root mean absolute error (RMAE). This reflects the differences between the predictions
by the model of the time-varying system at time t and the actual values of the data sets at
time t. The RMAE of the i-th individual at the g-th generation RMAE
i
g
can be calculated
based on (3).

( ) ( ) ( )
( )

=
D
N
j
D
D g
i
D
D
g
i
j t y
j t f j t y
N
RMAE
1
,
, , 1
% 100
x
, (3)
where
g
i
f is the polynomial model represented by the i-th individual
g
i
P at the g-th
generation, ( ) ( ) ( )
, , ,
D D
t j y t j x is the j-th data set at time t, and N
D
is the number of
training data sets used for developing the polynomial model of the time-varying system.
The velocity
,
g
i k
v (corresponding to the flight velocity in a search space) and the k-
th element of the i-th individual at the g-th generation
,
g
i k
p are calculated by expressions
(4) and (5) of the PSO [10] respectively:
( ) ( ) ( ) ( ) ( )
1
, 2
1
, , 1
1
, ,

+ + =
g
k i k
g
k i k i
g
k i
g
k i
p gbest rand p pbest rand v K v (4)
1
, , ,
g g g
i k i k i k
p p v

= + (5)
where

,1 ,2 ,
, , ...
p
i i i i N
pbest pbest pbest pbest
(
=

,

1 2
, , ...
p
N
gbest gbest gbest gbest
(
=

,
k = 1,2, , N
p
,
The best previous position so far of an individual is recorded from the previous
generation and is represented as
i
pbest ; the position of the best individual among all the
individuals is represented as gbest; rand() returns a uniform random number in the range
of [0,1]; w is an inertia weight factor;
1
and
2
are acceleration constants [13]; K is a
constriction factor derived from the stability analysis of equation (4) to ensure that the
system converges, but not prematurely. K is a function of
1
and
2
as reflected in the
following equation:

4 2
2
2

= K (6)
where
2 1
+ = and 4 > .
The proposed PSO utilizes pbest
i
and gbest to modify the current search point to
prevent the individuals from moving in the same direction, but to converge gradually
toward pbest
i
and gbest. g is the current generation number, G is the total number of
generations [14].
In (4), the particle velocity is limited by a maximum value
max
v . The parameter
max
v determines the resolution with which regions are to be searched between the present
position and the target position. This enhances the local exploration of a search process.
If
max
v is too high, individuals might fly past good solutions. If
max
v is too small,
individuals may not explore sufficiently beyond local solutions.
max
v was often set as
10%20% of the dynamic range of the element on each dimension. The pseudo code of
the proposed PSO is presented in Figure 1.

Figure 1: Pseudo code of the PSO

3. Polynomial modelling
In this section, the effectiveness of the PSO in modeling time-invariant or time-varying
systems is evaluated based on both the time-invariant data and time-varying data. The
PSO and the other commonly used, but recently developed, algorithms are compared.

3.1 Time-invariant and time-varying data
For time-invariant data, five sets of data, namely static data,
stat
Sph
H ,
stat
Ros
H ,
stat
Ras
H ,
stat
Gri
H and
stat
Ack
H , were generated based on each of the five benchmark functions, Sphere, Griewank,
Rastrigin, Rosenbrock and Ackley (see Table 3) by randomly choosing 100 numbers X
i
in
each of the predefined intervals | |
n
X X
max min
, . The 100 corresponding output responses Y
i

{
g0 // g is the generation number
Initialize a set of individuals
g
N
g g
pop
P P P ,..., ,
2 1
//
|
.
|

\
|
=
g
N i
g
i
g
i
g
i
p
p p p P
, 2 , 1 ,
,..., ,
Evaluate each individual
g
i
P based on (3)
while (g<G) do // G is the total number of generations
{
gg+1
Update the velocity
g
k i
v
,
based on (4)
if
g
k i
v
,
>v
max


g
k i
v
,
= v
max
end
if
g
k i
v
,
<v
max


g
k i
v
,
= v
max
end

Update each element of each particle
g
k i
p
,
based on (5)
Evaluate each particle
g
i
P based on (3)
Update pbest
i
and gbest
}
}
are computed by the benchmark function Y
i
= F(X
i
) whose landscape and optimum are
static with respect to time. The dimension of each benchmark function is n=4. The Sphere
(
Sph
F ) and Rosenbrock (
Ros
F ) functions are unimodal (a single local and global optimum),
and the Griewank (
Gri
F ), Rastrigin (
Ras
F ), and Ackley (
Ack
F ) functions are multimodal
(several local optima).

Table 3: Benchmark functions and initialization areas
Benchmark functions Initialization areas
| |
n
X X
max min
,
Sphere (Sph): ( ) ( )

=
=
n
i
i Sph
x F
1
2
x | |
n
50 , 50
Rosenbrock (Ros): ( ) ( ) ( ) ( )

=
+
+ =
1
1
2
2
2
1
1 100
n
i
i i i Ros
x x x F x | |
n
30 , 30
Rastrigin (Ras): ( ) ( ) ( )

=
+ =
n
i
i i Ras
x x F
1
2
10 2 cos 10 t x | |
n
12 . 5 , 12 . 5
Griewank (Gri): ( ) 1 cos
4000
1
1 1
2
+ |
.
|

\
|
=
[
= =
n
i
i
n
i
i Gri
i
x
x F x
| |
n
600 , 600
Ackley (Ack):
( )
( ) ( ) e x n
x n F
n
i
i
n
i
i Ack
+ +
|
.
|

\
|
=

=
=
20 2 cos 1 exp
1 2 . 0 exp 20
1
1
2
t
x

| |
n
32 , 32

The time-varying data used in this study was generated by a set of time-varying
functions which were extended from the benchmark functions shown in Table 3. X
i
was
generated by randomly chosen numbers in the predefined interval of the benchmark
function. Y
i
was computed in each generation of the PSO run by the time-varying
function Y
i
= F(X
i
, t) whose landscape or optima varies over time t. The mechanisms for
the development of the time-varying functions were based on the dynamic properties of
step changes of optima, changes of locations of optima, and changes of the landscapes of
the benchmark functions [33, 42].
For those based on the mechanism of step changes of optima, the time-varying
data was generated based on each of the five benchmark functions,
Sph
F ,
Ros
F ,
Ras
F ,
Gri
F
and
Ack
F . The optimum position
+
x of each benchmark function is moved by adding or
subtracting random values in all dimensions by a severity parameter s, at every change of
the environment [32]. The choice of whether to add or subtract the severity parameter s
on the optimum
+
x is done randomly with an equal probability. The severity parameter s
is defined by:
( ) ( )
( ) ( )

< A
> A +
=
0 rand if
0 rand if
min max
min max
X X d
X X d
s , (7)
where d A determines the scale of the step change of optima.
For each test run, a different random seed was used. The severity was chosen
relative to the extension (in one dimension) of the initialization area of each benchmark
function. The optima of the benchmark functions were periodically changed in every 100
generations of the runs of the algorithms. For small step changes of optima, 5% = Ad is
selected, and five sets of time-varying data, (namely step move data with % 5 = Ad ,
5
H
Step
Sph
,
5
H
Step
Ros
,
5
H
Step
Ras
,
5
H
Step
Gri
and
5
H
Step
Ack
) were generated based on the five benchmark
functions,
Sph
F ,
Ros
F ,
Ras
F ,
Gri
F and
Ack
F , in Table 3 respectively. For large step changes of
optima, 10% = Ad is selected, and five sets of time varying data, (namely step move
data with % 5 = Ad ,
10
H
Step
Sph
,
10
H
Step
Ros
,
10
H
Step
Ras
,
10
H
Step
Gri
and
10
H
Step
Ack
) were generated
respectively.
For those based on the mechanism of changes of locations of optima [21], the
time-varying data was generated based on the following time-varying function:
( ) ( ) ( ) ( ) ( ) ( ) s x F t w x F t w x F
i i
+ = 0 . 1 (8)
where ( )
|
.
|

\
|
=
100
floor
100
1
t
G
t w , G t s s 0 ; G is the pre-defined number of generations;
( ) x F
i
is any of the five benchmark functions,
Sph
F ,
Ros
F ,
Ras
F ,
Gri
F and
Ack
F , in Table 3, s
is a randomly chosen constant which is 20% of the range of the benchmark function
( ) x F
i
. The optimum in ( ) x F shifts from the original optimum
+
x of ( ) x F
i
to the new
optimum ( ) s +
+
x in every 100 generations. Based on these time-varying functions, five
sets of time varying data, namely shift data,
shift
Sph
H ,
shift
Ros
H ,
shift
Ras
H ,
shift
Gri
H and
shift
Ack
H , were
generated based on the five benchmark functions,
Sph
F ,
Ros
F ,
Ras
F ,
Gri
F and
Ack
F in Table 3
respectively.
For those based on the mechanism of changes of the landscapes of the benchmark
functions [24], the time-varying data was generated based on the following time-varying
function, which is similar to equation (8):
( ) ( ) ( ) ( ) ( ) ( ) x F t w s x F t w x F
j i
+ = 0 . 1 (9)
where ( ) x F
i
is any of the five benchmark functions in Table 3, and ( ) x F
j
is another
benchmark function. The landscape of ( ) x F changes gradually from the landscape of
( ) x F
j
to the landscape of ( ) x F
i
in every 100 generations. s is a randomly chosen constant
which is 20% of the range of the benchmark function ( ) x F
i
. Based on the time-varying
function (10), four sets of time varying data, namely match data,
match
Ros Sph
H ,
match
Ras Sph
H ,
match
Gri Sph
H
and
match
Ack Sph
H , were generated in which
Sph
F is used as ( ) x F
j
, and
Ros
F ,
Ras
F ,
Gri
F or
Ack
F
functions is used as ( ) x F
i
. Another three sets of match data,
match
Ras Ros
H ,
match
Gri Ros
H and
match
Ack Ros
H ,
were generated in which ( )
Ros
F x
is used as ( ) x F
j
, and
Ras
F ,
Gri
F or
Ack
F is used as ( ) x F
i
.
A brief summary of all the 27 data sets is presented in Table 4, and the benchmark
functions, which can be used to generate the data sets, can be downloaded from the
following link (http://www.4shared.com/account/dir/G2J--2eV/sharing.html).

Table 4: Description of the data sets
Data sets Descriptions
Static data
stat
Sph
H ,
stat
Ros
H ,
stat
Ras
H ,
stat
Gri
H ,
stat
Ack
H
The static data was generated by the benchmark
functions,
Sph
F ,
Ros
F ,
Ras
F ,
Gri
F and
Ack
F .
Step move
data with
% 5 = Ad
5
H
Step
Sph
,
5
H
Step
Ros
,
5
H
Step
Ras
,
5
H
Step
Gri
,
5
H
Step
Ack

The step move data with % 5 = Ad were generated
based on the benchmark functions,
Sph
F ,
Ros
F ,
Ras
F ,
Gri
F and
Ack
F , in which the locations of the optima
change from
+
x to s +
+
x in every 100 generations. s
is 5% of the ranges of the benchmark functions.
Step move
data with
% 10 = Ad
10
H
Step
Sph
,
10
H
Step
Ros
,
10
H
Step
Ras
,
10
H
Step
Gri
,
10
H
Step
Ack

The mechanism is the same as that for the above data
sets except that s is 10% of the ranges of the
benchmark functions.
Shift data
shift
Sph
H ,
shift
Ros
H ,
shift
Ras
H ,
shift
Gri
H ,
shift
Ack
H
The shift data was generated based on the benchmark
functions,
Sph
F ,
Ros
F ,
Ras
F ,
Gri
F and
Ack
F , in which the
locations of the optima move from
+
x to s +
+
x
gradually. s is 20% of the ranges of the benchmark
functions.
Match data
based on
Sph
F
match
Ros Sph
H ,
match
Ras Sph
H ,
match
Gri Sph
H ,
match
Ack Sph
H
The match data based on
Sph
F was generated in
which the landscape changes from
Sph
F to
Ros
F ,
Ras
F ,
Gri
F or
Ack
F .
Match data
based on
Ros
F
match
Ras Ros
H ,
match
Gri Ros
H ,
match
Ack Ros
H
The match data based on
Ros
F was generated in
which the landscape changes from
Ros
F to
Ras
F ,
Gri
F or
Ack
F .


3.2 Experiment Set-up
In this paper, because the basic operation of the PSO discussed in Section 2 is similar to
classical PSO, it is called classical PSO, C-PSO in this paper. The following parameters,
which can be found in reference [48], were implemented in the C-PSO: the number of
particles in the swarm was 100; the number of elements in the particle was 30; both the
acceleration constants
1
and
2
were set at 2.05; the maximum velocity
max
v was 0.2;
the pre-defined number of generations was 1000. Based on the results in [48], these
parameters can produce satisfactory results when solving both parameterized and
combinatorial problems. Therefore, these parameters are used in this research. The C-
PSO was compared against the following five approaches for generating models based on
both the time-invariant and time-varying data sets, which have been discussed in Section
3.1.
1. Classical genetic programming (C-GP): A commonly used method for
polynomial modeling, the classical genetic programming (C-GP) [25, 26] was
employed. Here the representation of the individuals of the grammatical
genetic programming [46] is identical to the one of the representations of the
C-PSO. The basic operations of the C-GP are shown in Figure 2 in the
Appendix. The C-GP first starts by creating a random initial population O(g)
of individuals [u
1
(g), u
2
(g), u
POP
(g)], while g=0. The i-th individual u
i
(g) at
the g-th generation represents the structure of the time-varying model (2). For
example, the i-th individual at the g-th generation u
i
(g) represents the structure
of the following time-varying model at time t:
u
i
(g) = x
1
2
x
2
2
+ x
1
x
2
x
4
(10)
After determining the structure of the time-varying model u
i
(g), the
system coefficients are determined. The completed time-varying model u
i
(g)
is represented by:
u
i
(g)= a
0
(t) + a
1
(t)x
1
2
a
2
(t)x
2
2
+ a
3
(t)x
1
x
2
x
4
(11)
where a
0
(t), a
1
(t), a
2
(t) and a
3
(t) are the system coefficients at time t, and are
calculated by OLSA. This is the same as the one used in the C-PSO for
calculating system coefficients. The classical genetic operations, point
mutation and one-point crossover, were used. Standard roulette wheel
selection was used. The following GA parameters were implemented in the C-
GP: The population size is 100. The type of replacement is elitist. Crossover
rate and mutation rate were 0.9 and 0.01 respectively. The pre-defined number
of generations was 1000. The dimension of the individuals was 30.
2. Dynamic particle swarm optimization (D-PSO): D-PSO is identical to the C-
PSO except for integration of the recent mechanisms for maintaining
diversities of the swarms [2] when solving the dynamic optimization problem.
The mechanism splits the whole set of particles into a set of interacting
swarms. These swarms interact locally through an exclusion parameter and
globally through an anti-convergence operator. Each swarm maintains its
diversity by using either charged or quantum particles. Results show that
when this mechanism for maintaining diversity in the PSO is used, the PSO
outperforms the other PSO or evolutionary algorithms, even where they are
integrated with other diversity maintaining mechanisms, for solving dynamic
optimization problems. The performance of D-PSO was optimized by tuning
it with different settings for the number particles in the sub-swarms. 5, 10 and
25 particles in the sub-swarms were used and the best performance among
them was recorded. The detailed description of the mechanisms used to
maintain diversity in the swarms can be found in [2].
3. Dynamic genetic programming (D-GP): D-GP is identical to the C-GP except
for integration of the recent mechanism [63] used for evolutionary algorithms
on solving dynamic optimization problems. The mechanism relocates the
positions of the individuals based on the changes of the landscape of the
dynamic optimization problem and the average sensitivities of their decision
variables to the corresponding change in the landscape. While integrating the
mechanism in the evolutionary algorithm, the evolutionary algorithm
outperforms the other dynamic evolutionary approaches for solving dynamic
optimization problems. The detailed description of the mechanisms for
maintaining diversity can be found in [63].
4. Polynomial-genetic algorithm (P-GA): P-GA is a genetic algorithm proposed by
Potgieter and Engelbrecht [53] which can evolve structurally polynomial
expressions in order to accurately describe a given data set. In P-GA, each
individual is used to represent the structure of the polynomial and this is
evolved based on the designed crossover and mutation operations. The
coefficients of the polynomial are determined by OLSA [3, 7]. The crossover
rate and the mutation rate were set at 0.1 and 0.2 respectively, which are the
same as those used in [54]. The population size was set at 100. The individual
length was set at 22.
5. Polynomial neural network (PNN): PNN is developed based on a genetic
algorithm which is proposed by Oh and Pedrycz [50]. Individuals in the
genetic algorithm are used to represent the parameters of the PNN including
the number of input variables, the order of the polynomial and input variables,
which lead to a structurally and parametrically optimized network. The
coefficients of the polynomial are determined by OLSA [3, 7]. The number of
layers of the PNN was set at 3. The crossover rate and the mutation rate used
in the genetic algorithm were set at 0.65 and 0.1 respectively, which are the
same as those used in [50]. The population size was set at 100. The individual
length was set at 36.

3.3 Experimental results
Thirty runs were performed on the C-PSO, D-PSO, D-GP, C-GP, P-GA and PNN in
generating polynomial models based on each of the 27 data sets shown in Table 4. In
each generation of the runs, the RMAE obtained by the individuals of the six algorithms
was recorded.
Online performance of the algorithms is demonstrated by the convergence plots.
Figures 3a, 3b, 3c, 3d and 3e show the convergence plots for the step move data with
% 5 = Ad (
5
H
Step
Sph
,
5
H
Step
Ros
,
5
H
Step
Ras
,
5
H
Step
Gri
,
5
H
Step
Ack
) respectively. It can be observed from
Figures 3a-3e that the evolutionary algorithms, C-GP, D-GP, P-GA and PNN, converged
more quickly in the early generations than did those of the PSO algorithms, C-PSO and
D-PSO. However, the PSO algorithms, C-PSO and D-PSO, kept progressing after the
early generations. Finally, both the PSO algorithms, C-PSO and D-PSO, reached a
smaller RMAE than that reached by the evolutionary algorithms, C-GP, D-GP, P-GA and
PNN, in the final stage of the search. D-PSO can reach the smallest RMAE compared
with those obtained by the other algorithms. Therefore in general, the PSO algorithms
outperform the evolutionary algorithms in generating the models for these static data sets
in later generations. For the rest of the data (static data, step move data with % 10 = Ad ,
shift data, match data based on
Sph
F , match data based on
Ros
F ), a similar finding can be
observed in that the convergence speed of the evolutionary algorithms was faster than
that of the particle swarm optimization algorithms in the early generations. In the late
generations, the particle swarm optimization algorithms can reach a smaller RMAE than
that reached by the evolutionary algorithms.
The smallest RMEA among all generations of each run of each algorithm was
recorded, and was averaged. This measure is called offline performance. The commonly
used method for testing the significance of the results, the Wilcoxon Rank Sum Test, was
used to compare the results between the two algorithms [19]. The results of the 30 runs
for two algorithms form two independent random samples X and Y. The distributions of X
and Y, F
X
and F
Y
, are compared using the null-hypothesis H
0
: F
X
=F
Y
and the one-sided
alternative H
1
: F
X
<F
Y
. We performed the significance tests at a significance level =
0.01. Only if the probability of the observed difference is less than , is the null-
hypothesis rejected and the alternative hypothesis accepted. The significance comparison
among a set of the six algorithms, C-PSO, D-PSO, D-GP, C-GP, P-GA and PNN, is
displayed using a 66 matrix, where X denotes that the result obtained by algorithm i
and that obtained by algorithm j is statistically significant different. i and j are the
position in the corresponding result table. An entry of _ indicates that the result
obtained by algorithm i and algorithm j is not a statistically significant difference. We
name such a matrix a significance matrix.
Tables 5, 6, 7, 8, 9 and 10 show the performance and the significance matrices for
the static data, the step move data with % 5 = Ad , the step move data with % 10 = Ad ,
the shift data, the match data based on
Sph
F and the match data based on
Ros
F respectively.
The average RMAE among the 30 runs of each algorithm and the ranks of the algorithms
in regard to the average RMAE are shown in the tables. Table 5 shows that D-PSO is
better than C-PSO in generating time-invariant models based on the static data
stat
Sph
H . C-
PSO is better than D-GP which is better than C-GP, P-GA and PNN. A significant
difference can be found between the results obtained by the PSO algorithms (C-PSO and
D-PSO) and those obtained by the evolutionary algorithms (D-GP, C-GP, P-GA and
PNN). However, there is no significant difference between the results obtained by C-PSO
and D-PSO, even if D-PSO can obtain a smaller RMAE than that obtained by C-PSO. In
regard to the other static data sets (
stat
Ros
H ,
stat
Ras
H ,
stat
Gri
H ,
stat
Ack
H ), both the PSO algorithms, C-
PSO and D-PSO can obtain a smaller average RMAE than that obtained by the
evolutionary algorithms, D-GP, C-GP, P-GA and PNN. Also, significant differences exist
between the results obtained by the PSO algorithms (C-PSO and D-PSO) and those
obtained by the evolutionary algorithms (D-GP, C-GP, P-GA and PNN). Also, similar
results can be found in Tables 6, 7, 8, 9 and 10 where the PSO algorithms are
significantly better than the evolutionary algorithms in generating models based on the
time-varying data.
Therefore, it can be concluded that the PSO algorithms are significantly better
than the evolutionary algorithms.
3.4 Population diversity
An investigation of population diversities of C-PSO, D-PSO, D-GP, C-GP, P-GA and
PNN is presented in this section. Maintaining population diversity in population-based
algorithms like evolutionary algorithms or PSO is a key to preventing premature
convergence and stagnation in local optima [11, 16, 40]. Thus it is essential to study the
population diversities of the six algorithms during the search. Various diversity measures,
which involve calculations of distance between two individuals in genetic programming
for the development of models, have been widely studied [4, 49]. These distance
measures calculate the distances between two individuals which are in a tree based
representation in genetic programming. They indicate the number of different nodes and
different terminals between two individuals. In this paper, we measure the distance
between two individuals by counting the number of different terms of the polynomials
represented by the two individuals in the four algorithms. If the terms in both
polynomials are all identical, the distance between the two polynomials is zero. The
distance between the two polynomials is larger when the number of different terms in the
two polynomials is larger. For example,
1
f and
2
f are two polynomials represented by:

5 3 1
2
4 3 1 2 1 1
x x x x x x x x f + + + + =
and
5 3 1 4 5 1 2 1 2
x x x x x x x x f + + + + =
Both
1
f and
2
f contain the three terms
1
x ,
2
x and
5 3 1
x x x , and the terms
3 1
x x
and
2
4
x in
1
f and the terms
5 1
x x and
4
x in
2
f are different. Therefore, the number of
terms which are different in
1
f and
2
f are 2, and the distance between
1
f and
2
f is
defined to be 2.
The diversity measure of the population at the g-th generation is defined by the
total sum of distances of individuals which is denoted as:
( ) ( ) ( )

= + =
=
p
N
i
p
N
i j
g g g
j s i s d
1 1
, o
where ( ) i s
g
and ( ) j s
g
are the i-th and the j-th individuals in the population at the g-th
generation, and d is the distance measure between the two individuals.
The diversities of the populations throughout the generations were recorded for
the four algorithms. Figure 4 shows the diversity plots which indicate the diversities of
the individuals in the algorithms in generating the models based on the step move data
with % 5 = Ad . Figure 4a, 4b, 4c, 4d and 4e shows the diversities for static data which
are generated based on the benchmark functions
Sph
F ,
Ros
F ,
Ras
F ,
Gri
F and
Ack
F respectively.
The diversities of the populations throughout the generations were recorded for the six
algorithms. The five figures indicate that the diversities along the generations of the D-
PSO are slightly higher than those of the C-PSO which are much higher than those of the
evolutionary algorithms, D-GP, C-GP, P-GA and PNN. For the rest of the data (static
data, step move data with % 10 = Ad , shift data, match data based on
Sph
F , match data
based on
Ros
F ), similar findings indicate that the diversities of the two PSO algorithms are
much larger than those of the evolutionary algorithms.
The findings indicate the reason why the PSO algorithms, D-PSO and C-PSO can
obtain significantly better results than the evolutionary algorithms, D-GP, C-GP, P-GA
and PNN. The diversities of the individuals of the PSO algorithms can be maintained
along the search in both earlier and later generations, while the individuals of the
evolutionary algorithms converged in the earlier generation. Therefore, the PSO
algorithms are more likely to explore the solution space, as the diversity of the
individuals of the algorithms can be maintained. In regard to the effectiveness of the two
PSO algorithms, since the diversities of the populations of C-PSO are only slightly
smaller than those of the D-PSO, D-PSO can obtain only slightly better results than those
obtained by C-PSO.

4 Conclusion and further work
In this paper, a particle swarm optimization algorithm has been proposed for developing
polynomial models in which both time-invariant and time-varying characteristics are
represented. The individuals of the PSO are represented by arithmetic operations and
system variables, which are the components of the polynomial models. A set of dynamic
benchmark functions whose optima or landscapes vary over time was employed to
evaluate the performance of the PSO. The PSO algorithms, C-PSO and D-PSO, were
used to generate models based on both the time-invariant and time-varying data sets. The
evolutionary algorithms, D-GP, C-GP, P-GA and PNN, were also included in the
experiments for comparison. Results show that the PSO algorithms significantly
outperform the evolutionary algorithms in generating models based on both the time-
invariant or time-varying data sets. It is observed that the evolutionary algorithms
converge faster in the earlier generations when compared with the PSO algorithms. In
contrast, the PSO algorithms can obtain better solutions in the later generations of the
runs. The performance obtained in the results can be explained by the diversity measures
and the fact that significant differences of diversities between the PSO algorithms and the
evolutionary algorithms exist.
In future work, we will enhance the effectiveness of the PSO by the hybridization
of the evolutionary algorithm and the PSO algorithm. Here the evolutionary algorithm
will be implemented to localize the potential solutions in the early generations and the
PSO algorithm will be implemented in order to continue to explore the solution space to
avoid pre-mature convergence in late generations. The resulting algorithm will be further
validated by solving real-time traffic flow forecasting problems, which are time varying
in nature.

References

[1] F.J. Berlanga, A.J. Rivera, M.L. del Jesus and F. Herrera, GP-COACH: Genetic
programming-based learning of COmpact and ACcurate fuzzy rule-based
classification systems for high-dimensional problems, Information Sciences, vol. 15,
no. 8, pp. 1183-1200, 2010.

[2] T. Blackwell and J. Branke, Multiswarms, exclusion, and anti-convergence in
dynamic environments, IEEE Transactions on Evolutionary Computation, vol. 10,
no. 4, pp. 459-472, 2006.

[3] S. Billings, M. Korenberg and S. Chen, Identification of nonlinear outputaffine
systems using an orthogonal least-squares algorithm, International Journal of
Systems Science, vol. 19, 1559-1568, 1988.

[4] E.K. Burke, S. Gustafson and G. Kendall, Diversity in genetic programming: an
analysis of measures and correlation with fitness, IEEE Transactions on
Evolutionary Computation, vol. 8, no. 1, pp. 47-62, 2004.

[5] Y.C. Chen, T.H.S. Li and Y.C. Yeh, EP-based kinematic control and adaptive fuzzy
sliding mode dynamic control for wheeled mobile robots, Information Sciences, vol.
179, no. 1, pp. 180-195, 2009.

[6] G. Chen and C.P. Low, Coordinating agents in shop floor environments from a
dynamic systems perspective, IEEE Transactions on Industrial Informatics, vol. 2,
no. 4, pp. 269-280, 2006.

[7] S. Chen, S. Billings and W. Luo, Orthogonal least squares methods and their
application to non-linear system identification, International Journal of Control, vol.
50, pp. 1873-1896, 1988.
[8] Y.P. Chen, W.C. Peng and M.C. Jian, Particle swarm optimization with
recombination and dynamic linkage discovery, IEEE Transactions on Systems, Man
and Cybernetics - Part B, vol. 37, no. 6, pp. 1460-1470, 2007.

[9] Y. Chen, B. Yang, Q. Meng and A. Abraham, Time series forecasting using a system
of ordinary differential equations, Information Sciences, vol. 181, no. 1, pp. 106-
114, 2011.

[10] M. Clerc and J. Kennedy, The particle swarm explosion, stability, and convergence
in a multidimensional complex space, IEEE Transactions on Evolutionary
Computation, vol. 6, no. 1, pp. 58-73, 2002.


[11] P. Day and A.K. Nandi, Robust text-independent speaker verification using genetic
programming, IEEE Transactions on Audio, Speech and Language Processing, vol.
15, no. 1, pp. 285-295, 2007.

[12] R. Eberhart and J. Kennedy, A new optimizer using particle swarm theory, in
Proceedings of the 6th International Symposium on Micro Machine and Human
Science, IEEE Service Center, Nagoya, Oct. 1995, pp.39-43.

[13] R.C. Eberhart and Y. Shi, Comparison between genetic algorithms and particle
swarm optimization, in Evolutionary Programming VII. New York: Springer-
Verlag, vol. 1447, Lecture Notes in Computer Science, 1998, pp. 611-616.

[14] R.C. Eberhart, and Y. Shi, Comparing inertia weights and constriction factors in
particle swarm optimization, in Proceedings of the Congress on Evolutionary
Computing, IEEE Service Center, vol. 1, Jul. 2000, pp.84-88.

[15] R.C. Eberhart and Y. Shi, Particle swarm optimization: developments, applications
and resources, in Proc. Congress on Evolutionary Computation (Hawaii), 2001, pp
816

[16] K.B. Edmund, G. Steven and K. Graham, Diversity in genetic programming: an
analysis of measures and correlation with fitness, IEEE Transactions on
Evolutionary Computation, vol. 8, no. 1, pp. 47-62, 2004.

[17] A. Ekart and S.Z. Nemeth, A metric for genetic programs and fitness sharing, in
Proceedings of the European Conference on Genetic Programming, vol. 1802, pp.
259-270, 2000.

[18] D. Gabor, W. Wides and R. Woodcock, A universal nonlinear filter predictor and
simulator which optimizes itself by a learning process, Proceedings of IEE, vol.
108-B, pp. 422-438, 1961.

[19] S. Garcia, D. Molina, M. Lozano and F. Herrera, A study on the use of non-
parametric tests for analyzing the evolutionary algorithms behaviour: a case study
on the CEC 2005 special session on real parameter optimization, Journal of
Heuristics, vol. 15, pp. 617-644, 2009.

[20] D.E. Goldberg, Genetic algorithms in search, optimization and machine learning,
United States of America: Addison Wesley Longman, Inc., 1989.

[21] H. Iba, Inference of differential equation models by genetic programming,
Information Sciences, vol. 178, pp. 4453-4468, 2008.

[22] A. Izadian, P. Khayyer and P. Famouri, Fault diagnosis of time-varying parameter
systems with application in MEMS LCRs, IEEE Transactions on Industrial
Electronics, vol. 56, no. 4, pp. 973-978, 2009.

[23] S. Janson and M. Middendorf, A hierarchical particle swarm optimizer for noisy and
dynamic environments, Genetic Programming and Evolvable Machines, vol. 7, pp.
329-354, 2006.

[24] Y. Jin and B. Sendhoff, Constructing dynamic optimization test problems using the
multi-objective optimization concept, In G. R. Raidl, editor, Applications of
evolutionary computing, Lecture Notes on Computer Sciences 3005, pp. 525-536.
Springer, 2004.

[25] J. Koza, Genetic Programming: On the Programming of Computers by Means of
Natural Evolution, MIT Press: Cambridge, 1992.

[26] J. Koza, Genetic Programming II: automatic discovery of reusable programs, MIT
Press, 1994.

[27] J.R. Koza, M.J. Streeter and M.A. Keane, Routine high-return human-competitive
automated problem-solving by means of genetic programming, Information
Sciences, vol. 178, no. 23, pp. 4434-4452, 2008.

[28] J. Kennedy and R.C. Eberhart, Particle Swarm Optimization, in IEEE International
Conference on Neural Networks, 1995, pp. 1942-1948.

[29] J. Kennedy and R.C. Eberhart, Swarm Intelligence. Morgan Kaufmann Publishers,
2001.

[30] J. Kennedy and R.C. Eberhart, A discrete binary version of the particle swarm
algorithm, IEEE Proceedings of International Conference on Systems, Man and
Cybernetics, vol. 5, pp. 4104-4108, 1997.

[31] C.J. Liao, C.T. Tseng and P. Luran, A discrete version of particle swarm
optimization for flowshop scheduling problems, Computer and Operations Research,
vol. 34, pp. 3099-3111, 2007.

[32] C. Li and S. Yang, A generalized approach to construct benchmark problems for
dynamic optimization, Proceedings of the 7
th
International Conference on Simulated
Evolution and Learning, pp. 391-400, 2008.

[33] C. Li, S. Yang, T.T. Nguyen, E.L. Yu, X. Yao, Y. Jin, H.G. Beyer and P.N.
Suganthan, Benchmark Generator for CEC 2009 Competition on Dynamic
Optimization, Technical report, University of Leicester and University of
Birmingham, UK, 2009.

[34] D.R. Liu, C.H. Lai and W.J. Lee, A hybrid of sequential rules and collaborative
filtering for product recommendation, Information Sciences, vol. 179, no. 20, pp.
3505-3519, 2009.

[35] G. Li, J.F. Wang, K.H. Lee and K.S. Leung, Instruction-matrix-based genetic
programming, IEEE Transactions on Systems, Man and Cybernetics Part B:
Cybernetics, vol. 38, no. 4, pp. 1036-1049, 2008.

[36] T. Li, Lei Guo and X. Lin, Improved delay-dependent bounded real lemma for
uncertain time-delay systems, Information Sciences, vol. 179, no. 20, pp. 3711-
3729, 2009.

[37] Y. Lin and B. Bhanu, Object detection via feature synthesis using MDL-based
genetic programming, IEEE Transactions on Systems, Man and Cybernetics Part
B: Cybernetics, vol. 35, no. 3, pp. 538-547, 2005.

[38] J. Madar, J. Abonyi and F. Szeifert, Genetic programming for the identification of
nonlinear input output models, Industrial and Engineering Chemistry Research, vol.
44, pp. 3178 3186, 2005.

[39] B. McKay, M.J. Willis and G.W. Barton, Steady-state modeling of chemical
processes using genetic programming, Computers and Chemical Engineering, vol.
21, no. 9, pp. 981-996, 1997.

[40] R.I. McKay, Fitness sharing genetic programming, Proceedings of the Genetic and
Evolutionary Computation Conference, pp. 435-442, 2000.

[41] N. Motoi, M. Ikebe and K. Ohnishi, Real-time gait planning for pushing motion of
humanoid robot, IEEE Transactions on Industrial Informatics, vol. 3, no. 2, pp. 154-
163, 2007.

[42] R.W. Morrison and K.A. De Jong, A test problem generator for non-stationary
environments, Proceedings of the IEEE Congress on Evolutionary Computation, pp.
2047-2053, 1999.

[43] D.P. Muni, N.R. Pal and J. Das, Genetic programming for simultaneous feature
selection and classifier design, IEEE Transactions on Systems, Man and
Cybernetics Part B: Cybernetics, vol. 36, no. 1, pp. 106-117, 2006.

[44] K. Natori, R. Oboe and K. Ohnishi, Stability analysis and practical design procedure
of time delayed control systems with communication disturbance observer, IEEE
Transactions on Industrial Informatics, vol. 4, no. 3, pp. 185-197, 2007.

[45] B. Naudts and L. Kallel, A comparison of predictive measures of problem difficulty
in evolutionary algorithms, IEEE Transactions on Evolutionary Computation, vol. 4,
no. 1, 2000.

[46] M.O. Neill and Conor Ryan, Grammatical Evolution, IEEE Transactions on
Evolutionary Computation, vol. 5, no. 4, pp. 349-358, 2001.

[47] M.O. Neill and A. Brabazon, Grammatical Swarm, Genetic and Evolutionary
Computation Conference, vol. 1, pp. 163-174, 2004.

[48] M.O. Neill and A. Brabazon, Grammatical Swarm: The generation of programs by
social programming, Natural Computing, vol. 5, pp. 443-462, 2006.

[49] X.H. Nguyen, R.I. McKay, D. Essam and H.A. Abbass, Toward an alternative
comparison between different genetic programming systems, Proceedings of
European Conference on Genetic Programming, pp. 67-77, 2004.

[50] S.K. Oh and W. Pedrycz, Multi-layer self-organizing polynomial neural networks
and their development with the use of genetic algorithms, Journal of the Franklin
Institute, vol. 343, pp. 125-136, 2006.

[51] Q.K. Pan, M.F. Tasgetiren and Y.C. Liang, A discrete particle swarm optimization
algorithm for the no-wait flowshop scheduling problem, Computers and Operations
Research, vol. 35, pp. 2907-2839, 2008.

[52] D. Parrott and X. Li, Locating and tracking multiple dynamic optima by a particle
swarm model using speciation, IEEE Transactions on Evolutionary Computation,
vol. 10, no. 4, pp. 440-458, 2006.

[53] K.E. Parsopoulos and M.N. Vrahatis, On the computation of all global minimizers
through particle swarm optimization, IEEE Transactions on Evolutionary
Computation, vol. 8, no. 3, pp. 211-224, 2004.

[54] G. Potgieter and A.P. Engelbrecht, Genetic algorithms for the structural optimization
of learned polynomial expression, Applied Mathematics and Computation, vol.
186, pp. 1441-1466, 2007.

[55] C.R. Reeves, Genetic algorithms and neighbourhood search, Evolutionary
Computing: AISB Workshop, pp. 115-130, 1994.

[56] K. Rodriguez-Vazquez, C.M. Fonseca and P.J. Fleming, Identifying the structure of
nonlinear dynamic systems using multiobjective genetic programming, IEEE
Transactions on Systems, Man and Cybernetics Part A, vol. 34, no. 4, pp. 531-545,
2004.

[57] T. Shibata and T. Murakami, Null space motion control by PID control considering
passivity in redundant manipulator, IEEE Transactions on Industrial Informatics,
vol. 4, no. 4, pp. 261-270, 2008.

[58] C.T. Tseng and C.J. Liao, A discrete particle swarm optimization for lot-streaming
flowshop scheduling problem, European Journal of Operational Research, vol. 191,
pp. 360-373, 2008.

[59] R.J.M. Vaessens, E.H.L. Aarts, and J.K. Lenstra, A local search template, in Proc
Parallel Problem-Solving from Nature 2, 1992, pp. 65-74.

[60] N. Wagner, Z. Michalewicz, M. Khouja and R.R. McGregor, Time series forecasting
for dynamic environments: the DyFor genetic program model, IEEE Transactions
on Evolutionary Computation, vol. 11, no. 4, pp. 433-452, 2007.

[61] J. Wang, Z. Fan, J.P. Terpenny and E.D. Goodman, Knowledge interaction with
genetic programming in mechatronic systems design using bond graphs, IEEE
Transactions on Systems, Man and Cybernetics Part C: Applications and Reviews,
vol. 35, no. 2, pp. 172-182, 2005.

[62] Y. Wang and Y. Yang, Particle swarm optimization with preference order ranking
for multi-objective optimization, Information Sciences, vol. 179, no. 12, pp. 1944-
1959, 2009.

[63] Y.G. Woldesenbet and G.G. Yen, Dynamic evolutionary algorithm with variable
relocation, IEEE Transactions on Evolutionary Computation, vol. 13, no. 3, pp.
500-513, 2009.

[64] S.X. Yang and X. Yao, Population-based incremental learning with associative
memory for dynamic environments, IEEE Transactions on Evolutionary
Computation, vol. 12, no. 5, pp. 542-561, 2008.

[65] A. Zafra and S. Ventura, G3P-MI: a genetic programming algorithm for multiple
instance learning, Information Sciences, vol. 180, no. 23, pp. 4496-4513, 2010.

[66] Y. Zhang, K.M. Jan, K.H. Ju and K.H. Chon, Representation of time-varying
nonlinear systems with time-varying principle dynamic modes, IEEE Transactions
on Biomedical Engineering, vol. 54, no. 11, pp. 1983-1992, 2007.








Appendix


Figure 2: The pseudocode of the genetic programming GP

{
Step 1: g=0
Step 2: Initialize O(g)=[u
1
(g), u
2
(g), u
POP
(g)]
// O(g) is the population of the g-th generation.
// u
i
(g) is the i-th individual of O(g).
//where u
k
(g) = x
i
+ x
i
x
j
+..[x
i

Step 3: Assign system coefficients a(t) in all u
k
(g) by LSM
//where u
i
(g) = a
0
(t) + a
i
(t) x
i
+ a
ij
(t) x
i
x
j
+
// ..+ a
12..Nterm
(t) [x
i

Step 4: Evaluate all u
k
(g) based on (3)
while (Terminational condition not fulfilled) do {
Step 5: Parent Selection O(g+1) =[u
1
(g+1), u
2
(g+1),
u
POP
(g+1)]
// where u
k
(g+1) = x
i
+ x
i
x
j
+[x
i

Step 6: Crossover O(g+1)
Step 7: Mutation O(g+1)
Step 8: Assign parameters a(k) in all u
k
(g+1) by
LSM
//where u
k
(g+1) = a
0
(t) + a
i
(t) x
i
+ a
ij
(t) x
i
x
j
+
// ..+ a
12..Nterm
(t) [x
i

Step 9: Evaluate all u
k
(g+1) based on (3)
Step 10: O(g)= O(g+1)
Step 11: g=g+1
}
0 100 200 300 400 500 600 700 800 900 1000
10
1
10
2
generation number
m
e
a
n

a
b
s
o
l
u
t
e

e
r
r
o
r
Sphere (5%)


C-PSO
D-PSO
D-GP
C-GP
P-GA
PNN

Figure 3a: Convergence plot for step moving data
5
H
Step
Sph
(Sphere function with 5% = Ad )
0 100 200 300 400 500 600 700 800 900 1000
10
2
10
3
10
4
generation number
m
e
a
n

a
b
s
o
l
u
t
e

e
r
r
o
r
Rosenbrock (5%)


C-PSO
D-PSO
D-GP
C-GP
P-GA
PNN

Figure 3b: Convergence plot for step moving data
5
H
Step
Ros
(Rosenbrock function with
5% = Ad )
0 100 200 300 400 500 600 700 800 900 1000
10
1
10
2
generation number
m
e
a
n

a
b
s
o
l
u
t
e

e
r
r
o
r
Rastrigrin (5%)


C-PSO
D-PSO
D-GP
C-GP
P-GA
PNN

Figure 3c: Convergence plot for step moving data
5
H
Step
Ras
(Rastrigrin function with
5% = Ad )
0 100 200 300 400 500 600 700 800 900 1000
10
1
10
2
generation number
m
e
a
n

a
b
s
o
l
u
t
e

e
r
r
o
r
Griewank (5%)


C-PSO
D-PSO
D-GP
C-GP
P-GA
PNN

Figure 3d: Convergence plot for step moving data
5
H
Step
Gri
(Griewank function with
5% = Ad )
0 100 200 300 400 500 600 700 800 900 1000
10
0
10
1
generation number
m
e
a
n

a
b
s
o
l
u
t
e

e
r
r
o
r
Ackley (5%)


C-PSO
D-PSO
D-GP
C-GP
P-GA
PNN

Figure 3e: Convergence plot for step moving data
5
H
Step
Ackley
(Ackley function with 5% = Ad )





0 100 200 300 400 500 600 700 800 900 1000
10
3
10
4
10
5
10
6
generation number
D
i
f
f
e
r
e
n
c
e

i
n

p
o
p
u
l
a
t
i
o
n
Sphere(5%)


C-PSO
D-PSO
D-GP
C-GP
P-GA
PNN

Figure 4a: Diversity plot for step moving data
5
H
Step
Sph
(Sphere function with 5% = Ad )
0 100 200 300 400 500 600 700 800 900 1000
10
3
10
4
10
5
10
6
generation number
D
i
f
f
e
r
e
n
c
e

i
n

p
o
p
u
l
a
t
i
o
n
Rosenbrock(5%)


C-PSO
D-PSO
D-GP
C-GP
P-GA
PNN

Figure 4b: Diversity plot for step moving data
5
H
Step
Ros
(Rosenbrock function with
5% = Ad )

0 100 200 300 400 500 600 700 800 900 1000
10
3
10
4
10
5
10
6
generation number
D
i
f
f
e
r
e
n
c
e

i
n

p
o
p
u
l
a
t
i
o
n
Rastrigrin(5%)


C-PSO
D-PSO
D-GP
C-GP
P-GA
PNN

Figure 4c: Diversity plot for step moving data
5
H
Step
Ras
(Rastrigrin function with 5% = Ad )
0 100 200 300 400 500 600 700 800 900 1000
10
3
10
4
10
5
10
6
generation number
D
i
f
f
e
r
e
n
c
e

i
n

p
o
p
u
l
a
t
i
o
n
Griewank(5%)


C-PSO
D-PSO
D-GP
C-GP
P-GA
PNN

Figure 4d: Diversity plot for step moving data
5
H
Step
Gri
(Griewank function with 5% = Ad )
0 100 200 300 400 500 600 700 800 900 1000
10
3
10
4
10
5
10
6
generation number
D
i
f
f
e
r
e
n
c
e

i
n

p
o
p
u
l
a
t
i
o
n
Ackley(5%)


C-PSO
D-PSO
D-GP
C-GP
P-GA
PNN

Figure 4e: Diversity plot for step moving data
5
H
Step
Ackley
(Ackley function with 5% = Ad )




Table 5: Average RMAE and ranks for the static data (
stat
Sph
H ,
stat
Ros
H ,
stat
Ras
H ,
stat
Gri
H ,
stat
Ack
H )
Algorithm
Static
Sph
F
(
stat
Sph
H )
Static
Ros
F
(
stat
Ros
H )
Static
Ras
F
(
stat
Ras
H )
Static
Gri
F
(
stat
Gri
H )
Static
Ack
F
(
stat
Ack
H )
Mean
rA
oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
1 C-PSO 0.8873
(1)
_ _ X X X X 1.9252
(2)
_ _ X X X X 3.6538
(2)
_ _ X X X X 2.9008
(2)
_ _ X X X X 0.8178
(2)
_ X X X X X 1.8
2 D-PSO 1.3661
(2)
_ _ X X X X 1.0401
(1)
_ _ X X X X 2.2758
(1)
_ _ X X X X 1.8290
(1)
_ _ X X X X 0.7747
(1)
X _ X X X X 1.2
3 D-GP 4.3595
(3)
X X _ X _ _ 4.2012
(3)
X X _ _ _ X 5.6365
(3)
X X _ _ _ X 3.4597
(3)
X X _ _ _ _ 0.9546
(3)
X X _ X X X 3
4 C-GP 5.7800
(6)
X X X _ _ _ 5.2979
(5)
X X _ _ _ _ 6.0036
(4)
X X _ _ _ _ 5.8344
(6)
X X _ _ _ _ 1.0398
(5)
X X X _ _ _ 5.2
5 P-GA 4.7552
(5)
X X _ _ _ _ 5.1909
(4)
X X _ _ _ _ 6.0764
(5)
X X _ _ _ _ 5.1195
(5)
X X _ _ _ _ 1.0069
(4)
X X X _ _ _ 4.6
6 PNN 4.4885
(4)
X X _ _ _ _ 5.8052
(6)
X X X _ _ _ 6.3742
(6)
X X X _ _ _ 4.9114
(4)
X X _ _ _ _ 1.1207
(6)
X X X _ _ _ 5.2

oP Average RMAE obtained by the algorithms; rA rank; X difference between the two algorithms is significant; _ -
difference between the two algorithms is not significant;














Table 6: Average RMAE and ranks for the step move data with % 5 = Ad (
5
H
Step
Sph
,
5
H
Step
Ros
,
5
H
Step
Ras
,
5
H
Step
Gri
,
5
H
Step
Ack
)
Algorithm
Step
Sph
F - % 5 = Ad
(
5
H
Step
Sph
)
Step
Ros
F - % 5 = Ad
(
5
H
Step
Ros
)
Step
Ras
F - % 5 = Ad
(
5
H
Step
Ras
)
Step
Gri
F - % 5 = Ad
(
5
H
Step
Gri
)
Step
Ack
F - % 5 = Ad
(
5
H
Step
Ack
)
Mean
rA
oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
1 C-PSO 2.8118
(2)
_ _ X X X X 2.0365
(2)
_ _ X X X X 4.9994
(2)
_ X X X X X 4.4084
10
1
(2)
_ X X X X X 0.8790
(2)
_ _ X X X X 2
2 D-PSO 1.7906
(1)
_ _ X X X X 1.6411
(1)
_ _ X X X X 4.9159
(1)
X _ X X X X 2.6067
10
1
(1)
X _ X X X X 0.7767
(1)
_ _ X X X X 1
3 D-GP 4.1972
(3)
X X _ X _ _ 4.8945
(3)
X X _ _ _ _ 6.6285
(5)
X X _ _ _ _ 6.0471
10
1
(3)
X X _ _ _ _ 2.7233
(4)
X X _ _ _ _ 3.6
4 C-GP 6.2698
(6)
X X X _ _ _ 6.0860
(6)
X X _ _ _ _ 6.7806
(6)
X X _ _ _ _ 8.3000
10
1
(6)
X X _ _ _ _ 3.0748
(6)
X X _ _ _ _ 6
5 P-GA 5.9230
(5)
X X _ _ _ _ 5.2762
(4)
X X _ _ _ _ 6.2917
(3)
X X _ _ _ _ 6.3092
10
1
(4)
X X _ _ _ _ 2.9372
(5)
X X _ _ _ _ 4.2
6 PNN 5.1709
(4)
X X _ _ _ _ 5.7928
(5)
X X _ _ _ _ 6.5860
(4)
X X _ _ _ _ 6.7838
10
1

(5)
X X _ _ _ _ 2.7220
(3)
X X _ _ _ _ 4.2
oP Average RMAE obtained by the algorithms; rA rank; X difference between the two algorithms is significant; _ -
difference between the two algorithms is not significant;











Table 7: Average RMAE and ranks for the step move data sets with % 10 = Ad (
10
H
Step
Sph
,
10
H
Step
Ros
,
10
H
Step
Ras
,
10
H
Step
Gri
,
10
H
Step
Ack
)
Alg.
Step
Sph
F - % 10 = Ad
(
10
H
Step
Sph
)
Step
Ros
F - % 10 = Ad
(
10
H
Step
Sph
)
Step
Ras
F - % 10 = Ad
(
10
H
Step
Sph
)
Step
Gri
F - % 10 = Ad
(
10
H
Step
Sph
)
Step
Ack
F - % 10 = Ad
(
10
H
Step
Sph
)
Mean
rA
oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
1 C-
PSO
3.8595
(2)
_ _ X X X X 4.1047
(2)
_ X X X X X 5.6531
(2)
_ _ X X X X 6.7125
10
1
(2)
_ X X X X X 6.1591
10
1
(2)
_ _ X X X X 2
2 D-
PSO
3.6369
(1)
_ _ X X X X 3.1930
(1)
X _ X X X X 5.6379
(1)
_ _ X X X X 5.1045
10
1
(1)

X _ X X X X 3.5942
10
1
(1)
_ _ X X X X 1
3 D-
GP
5.1225
(5)
X X _ X _ _ 7.2864
(5)
X X _ _ _ _ 7.2221
(4)
X X _ _ _ _ 0.8528
10
2
(3)
X X _ X _ _ 0.7737
10
2
(3)
X X _ _ _ _ 4
4 C-
GP
7.1423
(6)
X X X _ _ _ 7.5441
(6)
X X _ _ _ _ 7.5313
(5)
X X _ _ _ _ 1.0793
10
2
(6)
X X X _ _ _ 1.0155
10
2
(6)
X X _ _ _ _ 5.8
5 P-
GA
5.9374
(4)
X X _ _ _ _ 7.0326
(3)
X X _ _ _ _ 7.1957
(3)
X X _ _ _ _ 1.0019
10
2
(5)
X X _ _ _ _ 0.9216
10
2
(5)
X X _ _ _ _ 4
6 PNN 5.3002
(3)
X X _ _ _ _ 7.1971
(4)
X X _ _ _ _ 7.9044
(6)
X X _ _ _ _ 0.8770
10
2

(4)
X X _ _ _ _ 0.8781
10
2
(4)
X X _ _ _ _ 4.2
oP Average RMAE obtained by the algorithms; rA rank; X difference between the two algorithms is significant; _ -
difference between the two algorithms is not significant;











Table 8: Average RMAE and ranks for the shift data (
shift
Sph
H ,
shift
Ros
H ,
shift
Ras
H ,
shift
Gri
H ,
shift
Ack
H )
Alg.
Shift
Sph
F
(
shift
Sph
H )
Shift
Ros
F
(
shift
Ros
H )
Shift
Ras
F
(
shift
Ras
H )
Shift
Gri
F
(
shift
Gri
H )
Shift
Ack
F
(
shift
Ack
H )
Mean
rA
oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
1 C-
PSO
8.0110
-3
(1)
_ _ X X X X 6.9010
-5
(2)
_ X X X X X 0.96
(2)
_ _ X X X X 2.4610
-1
(2)
_ X X X X X 3.5610
-1
(2)
_ _ X X X X 1.8
2 D-
PSO
8.5510
-3
(2)
_ _ X X X X 3.6010
-5
(1)
X _ X X X X 0.90
(1)
_ _ X X X X 0.4810
-1
(1)
X _ X X X X 3.2710
-1
(1)
_ _ X X X X 1.2
3 D-GP 15.3210
-3

(6)
X X _ _ _ _ 7.4510
-5
(4)
X X _ _ _ _ 2.11
(3)
X X _ _ _ _ 2.7310
-1
(3)
X X _ X _ _ 3.6610
-1
(3)
X X _ _ X _ 3.8
4 C-GP 12.9810
-3
(3)
X X _ _ _ _ 7.5810
-5
(5)
X X _ _ _ _ 3.09
(6)
X X _ _ _ _ 3.7110
-1
(6)
X X X _ _ _ 3.7410
-1
(4)
X X _ _ _ _ 4.8
5 P-GA 13.2610
-3

(4)
X X _ _ _ _ 7.2010
-5
(3)
X X _ _ _ _ 2.85
(5)
X X _ _ _ _ 3.4710
-1
(5)
X X _ _ _ _ 4.5710
-1
(6)
X X X _ _ _ 4.6
6 PNN 14.3110
-3

(5)
X X _ _ _ _ 7.8310
-5

(6)
X X _ _ _ _ 2.56
(4)
X X _ _ _ _ 3.3310
-1

(4)
X X _ _ _ _ 3.8510
-1

(5)
X X _ _ _ _ 4.8
oP Average RMAE obtained by the algorithms; rA rank; X difference between the two algorithms is significant; _ -
difference between the two algorithms is not significant;















Table 9: Average RMAE and ranks for the match data based on
Sph
F (
match
Ros Sph
H ,
match
Ras Sph
H ,
match
Gri Sph
H ,
match
Ack Sph
H )
Alg.
Match
Sph
F -
Ros
F
(
match
Ros Sph
H )
Match
Sph
F -
Ras
F
(
match
Ras Sph
H )
Match
Sph
F -
Gri
F
(
match
Gri Sph
H )
Match
Sph
F -
Ack
F
(
match
Ack Sph
H )
Mean
rA
oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
1 C-
PSO
2.382110
-4
(2)
_ _ X X X X 3.0658
(2)
_ X X X X X 2.9197
(2)
_ _ X X X X 1.077110
-1
(2)
_ X X X X X 2
2 D-
PSO
1.971310
-4
(1)
_ _ X X X X 0.7883
(1)
X _ X X X X 2.7702
(1)
_ _ X X X X 3.770610
-2
(1)
X _ X X X X 1
3 D-GP 3.273510
-3
(3)
X X _ _ _ _ 4.5844
(3)
X X _ _ _ _ 5.0644
(3)
X X _ X _ _ 2.005110
-1
(3)
X X _ X X X 3
4 C-GP 4.130310
-4
(5)
X X _ _ _ _ 4.6479
(4)
X X _ _ _ _ 6.0078
(6)

X X X _ _ _ 2.912810
-1
(6)
X X X X _ _ 5.25
5 P-GA 4.090110
-4
(4)
X X _ _ _ _ 5.2263
(6)
X X _ _ _ _ 5.7057
(5)
X X _ _ _ _ 2.832410
-1
(5)
X X X _ _ _ 5
6 PNN 4.451510
-4

(6)
X X _ _ _ _ 4.7533
(5)
X X _ _ _ _ 5.6600
(4)
X X _ _ _ _ 2.805710
-1

(4)
X X X _ _ _ 4.75
oP Average RMAE obtained by the algorithms; rA rank; X difference between the two algorithms is significant; _ -
difference between the two algorithms is not significant;















Table 10: Average RMAE and ranks for the match data based on
Ros
F (
match
Ras Ros
H ,
match
Gri Ros
H ,
match
Ack Ros
H )
Alg.
Match
Sph
F -
Ros
F
(
match
Ros Sph
H )
Match
Sph
F -
Ras
F
(
match
Ras Sph
H )
Match
Sph
F -
Gri
F
(
match
Gri Sph
H )
Mean
rA
oP
(rA)
S-matrix oP
(rA)
S-matrix oP
(rA)
S-matrix
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
1 C-
PSO
2.1106
(2)
_ X X X X X 4.3659
(2)
_ X X X X X 1.202210
-1
(2)
_ X X X X X 2
2 D-
PSO
1.5180
(1)
X _ X X X X 3.5196
(1)
X _ X X X X 4.886410
-2
(1)
X _ X X X X 1
3 D-GP 5.5616
(4)
X X _ _ _ _ 4.9281
(3)
X X _ _ _ _ 2.089910
-1
(3)
X X _ X _ X 3.33
4 C-GP 5.8036
(5)
X X _ _ _ _ 5.0353
(4)
X X _ _ _ _ 2.564810
-1
(5)
X X X _ _ _ 4.67
5 P-GA 5.6969
(6)
X X _ _ _ _ 5.1566
(5)
X X _ _ _ _ 2.545010
-1
(4)
X X _ _ _ _ 5
6 PNN 5.5449
(3)
X X _ _ _ _ 5.1725
(6)
X X _ _ _ _ 2.593510
-1

(6)
X X X _ _ _ 4
oP Average RMAE obtained by the algorithms; rA rank; X difference between the two algorithms is significant; _ -
difference between the two algorithms is not significant;